Andrew Reed

1. Key Themes

Human Click Data Is Actively Harmful for Agentic Search

The foundational insight of Parallel is that the feedback loop that made Google great is the wrong feedback loop for agentic search. Rather than using clicks to signal relevance, agents can generate their own ground-truth evaluations, unlocking a new paradigm for ranking and indexing.

"Our view at Parallel is that human click data is a bug and agent doing work with search should rely on agent feedback, not human feedback." 00:00:00

The Index as a Latency Optimization — and How to Skip It

Parag reframes the web index not as a fundamental necessity but as a latency optimization, which opened the door for Parallel to bootstrap without the full infrastructure investment upfront. By launching a search agent product first — allowing real-time crawling at query time — they could build their index incrementally while generating revenue.

"An index is oftentimes you can think of it as a latency optimization. So if you give up on that dimension, if you're competing with humans, that's why our search agents were competing with the alternative being outsourcing to humans to curate amazing data." 00:08:17

"By building a product that was a search agent to do real work on top of web data, we were able to incrementally go build our index." 00:08:42

The 1000x Multiplier: Agent Queries Will Dwarf Human Queries

Even conservative agentic use cases already generate 5–20x the number of searches a human would make. Background agents and orchestrated multi-agent workflows push this to 100–1000x or more per user per day — a fundamental demand shift that makes agentic search a category-defining infrastructure layer.

"If you run a typical search agent and even without doing deep research, it'll do somewhere between five to 20 searches, even if it answers within a few seconds... every time you write a prompt to it in ChatGPT, it will do five to 10 searches. As you dial it up, it'll do hundreds and thousands of searches." 00:28:07

"I bet they're doing thousand X more than that... maybe hundred to a thousand. If you count some of the things that happen at my company, which isn't like assigned to a human, it might easily be more than a thousand X." 00:33:05

The Old Internet Business Model Is Breaking

The ad-supported web was built on human attention scarcity. Agents don't click ads, don't convert into subscriptions in trackable ways, and don't reward the content creators whose data they consume. This creates an existential misalignment between how the web monetizes today and how it will be consumed tomorrow.

"If we don't figure out a new business model we're seeing it already. Like people are going to say okay I don't want my content to be accessed by an agent because I have a business model for humans." 00:36:37

"Your one available business model pre-Parallel was to be in the head and be able to transact with the lab on some fixed fee contract which includes training and liability and then inference time access. That option is not available to most content on the web." 00:37:37

Shapley Values as the Economic Foundation for the Agentic Web

Parallel's answer to content monetization is a game-theoretic framework: Shapley values measure a piece of content's counterfactual contribution to agent output quality. This enables differential pricing that rewards high-quality, unique content and high-value use cases — creating incentive alignment that could sustain the content ecosystem.

"You have to ask the question of how much incremental value did somebody's content add... The formalization of this kind of an intuition is the core framework we use. It's called Shapley values." 00:40:58

"If you have unique differentiated data you get paid more. If a banker in an expensive job reads your data versus my retired dad reads your data, the banker ends up paying more for that read because it's part of high value work." 00:46:00

Quality, Cost, Latency — and the Disciplined Sequencing to Get There

Parallel explicitly chose to ignore latency for the first two years, focusing only on quality and cost. Once those were nailed, they attacked latency — shipping from a 3-second response to 200 milliseconds recently with a product called TurboNow. This sequencing discipline is a deliberate strategic choice.

"For the first couple of years, we said, let's focus, let's ignore latency and let's just nail the other two. Because optimizing systems, distilling to smaller models is much more known art than unknown research." 00:31:50

"Last week, we shipped a product which now does it in 200 milliseconds." 00:12:16

The Web Is Shifting from Pull to Push

The final phase of Parag's vision is that agents stop reactively querying the web and instead register standing interests — the web notifies agents when something relevant changes. This transforms search from a point-in-time tool into a continuous intelligence feed.

"The web goes from pull to push... not a point in time need but a long term — here is what is actionable for me if something like this happens as evidenced by all of the information on the web that is changing all the time, call me, and then I'll run my agent on it." 00:52:02

Agents Are Now a Distinct Customer Class, Not Just a Technology

Parag explicitly distinguishes agents as a new customer segment, not just a technical capability. This reframing changes how you build products, price them, and design business models around them.

"It's way more tractable as a problem because of the existence of agents, not just as in technology, but as a distinct customer." 00:00:24

2. Contrarian Perspectives

Human Click Data — the Foundation of Google's Moat — Is a Bug

Most people assume Google's click-through data is an insurmountable competitive advantage. Parag argues the opposite: it's the wrong signal for the next era of search.

"Our view at Parallel is that human click data is a bug. An agent doing work with search should rely on agent feedback, not human feedback." 00:00:00

SEO Slop Exists Because Humans Are Lazy, Not Because Publishers Are Bad Actors

Parag reframes what most people call content pollution as a rational, value-adding adaptation to human laziness and browser limitations. The problem isn't bad publishers — it's the interface mismatch between humans and authoritative data.

"You can call it slop, pre-AI human slop. Or you can call it catering to a lazy human and being successful at SEO... with agents, we're not making the agent click around and fumble around and grab a PDF." 00:18:41

Fixed-Price Training Data Deals with AI Labs Are a Broken Business Model — Even for Big Publishers

The conventional wisdom is that striking a deal with OpenAI or Anthropic is a windfall for content owners. Parag argues those deals structurally decline in value as AI inference scales massively, making them unsustainable even for the largest publishers.

"When AI's inference grows let's say 7x this year and another 7x the next year on this 50x, their deal size is not growing 50x. Like none of them after signing a two year deal believes that their share isn't going to decline materially at renewal." 00:38:06

Ads Are Not the Villain — They Are the Most Efficient Differential Pricing Mechanism Ever Built

Against the popular narrative that ads are exploitative, Parag argues ads are actually the most economically elegant solution to differential pricing ever deployed at scale — and that their breakdown in the agent era is a genuine loss, not a liberation.

"Ads is extraordinarily efficient at differential pricing and monetization on the web which is why it has been a dominant business model." 00:36:07

A Complement to a Model Is Harder to Displace Than a Model Itself

While the conventional concern is that every model improvement displaces infrastructure layers, Parag inverts this: a better model expands Parallel's addressable use cases rather than threatening them.

"Somebody ships a better model, they have now unlocked four more use cases where we can be valuable." 00:11:21

3. Companies Identified

Parallel (Web Systems)

Agentic search infrastructure company building a dedicated web index, ranking models, and content monetization layer for AI agents. Founded by Parag Agrawal (former Twitter CEO). Mentioned as the central company being discussed throughout. Recently launched TurboNow, a 200ms agentic search product, and announced a Google Cloud partnership as a search and grounding provider for enterprise agent APIs.

"At Parallel, we're building a bunch of technology in order to allow agents to search and use the web... you started Parallel with the bet that agents would do it a thousand X more than humans ever have." 00:01:25

Google Cloud

Hyperscaler and AI platform. Mentioned as having formally partnered with Parallel as a search and grounding provider for enterprise agent APIs built on GCP, specifically for Gemini model grounding.

"We announced today, actually, that we are working with Google Cloud to be a search and grounding provider for their enterprise agent APIs. So if you think of grounding Gemini models or other models available on GCP, when you build agents on GCP... your options are Google search or parallel search." 00:25:51

Cloudflare

Network infrastructure and security company. Mentioned for a notable data point: their monitoring shows AI web traffic is now roughly equivalent to human traffic in terms of page reads.

"I recently saw, I think it was Cloudflare that said that in their monitoring of web traffic, the AI traffic is about the same as human traffic in terms of page reads." 00:33:35

Notion

Productivity and collaboration platform. Mentioned as a concrete example of an agent orchestration platform Parag personally uses to build custom meeting prep agents powered by Parallel's APIs.

"I use Notion's agent and you can build custom agents which look into all the internal data that we have at Parallel plus all of the web data using Parallel's APIs to create meeting prep docs." 00:30:36

OpenAI / ChatGPT

AI lab and consumer AI product. Mentioned as an example of an agentic product that already fires 5–10 searches per user prompt, and as a party to content licensing deals that Parag sold Twitter data to.

"Every time you write a prompt to it in ChatGPT, it will do five to 10 searches." 00:28:35

Twitter / X

Social media platform. Mentioned in the context of Parag's prior CEO role, ads business model lessons, and having transacted data deals with OpenAI.

"I sold Twitter data to OpenAI, having transacted on that side to figure out what actually might work." 00:39:06

Stack Overflow

Developer Q&A platform. Mentioned as a concrete example of a human-traffic-dependent business being devastated by the shift to agent-mediated information access.

"You see the traffic data, the Stack Overflow plummeting. You see a lot of the human internet as we know it going away because of incentives." 00:47:26

SEC (U.S. Securities and Exchange Commission)

Regulatory body. Mentioned to illustrate how authoritative sources (SEC filings) are bypassed by humans in favor of derivative, SEO-optimized pages — a problem agents can solve by going directly to the source.

"If you ask for a public company's most recent financials, we can sit here and know that there exists an authoritative filing with the SEC, which has that number. Perhaps on page 73 of a PDF." 00:16:56

4. People Identified

Parag Agrawal

Former CTO and CEO of Twitter; founder and CEO of Parallel Web Systems. Mentioned throughout as the primary guest. Notable for having sold Twitter data to OpenAI, for founding Parallel originally as "Shapley Inc.," and for having deep technical expertise in web-scale systems, search, ads, and AI infrastructure.

"Parag Agrawal of Twitter CEO fame was the CEO of Twitter before selling it to Elon and is now back on the founder arc. You founded a company called Parallel Web Systems, which is scaling up agentic search for the agentic web." 00:00:57

Partner at Sequoia Capital. Co-host on this episode of Training Data. Mentioned for making his podcast debut and for sharp questions on the infrastructure vs. AI research nature of Parallel.

"I'm Andrew. I didn't get an introduction, but I'm also happy to be an inaugural guest on the Training Data podcast." 00:03:14

Sonya Huang

Partner at Sequoia Capital. Host of Training Data podcast. Asks the driving technical and strategic questions throughout the episode.

"Thank you, Sonya, for having me." 00:01:25

James Flynn

Identified as a "great young guy" at Sequoia Capital, apparently building or running an internal agent for meeting prep documents.

"I was going to say James Flynn, but he's one of our great young guys." 00:30:30

Elon Musk

Referenced briefly in context of acquiring Twitter from Parag.

"Parag Agrawal of Twitter CEO fame was the CEO of Twitter before selling it to Elon." 00:00:57

Lloyd Shapley (implied)

The mathematician behind Shapley values, a game-theoretic concept that Parallel uses as the core framework for content attribution and compensation.

"Shapley value is this very theoretical mathematical concept... the three of us collaborate on something and the whole is bigger than the sum of parts — in that moment the theoretical question, mathematical way of effectively answering this question." 00:42:31

5. Operating Insights

Sequence Your Optimization Dimensions Deliberately — Don't Optimize Everything at Once

Parallel explicitly chose to treat quality and cost as the first two years' sole focus, deferring latency until those were solved. This allowed the team to tackle what was "more known art than unknown research" first, building confidence before attacking the hardest systems problem. This is a replicable framework for any company with multiple competing optimization axes.

"For the first couple of years, we said, let's focus, let's ignore latency and let's just nail the other two. Because optimizing systems, distilling to smaller models is much more known art than unknown research... Once we achieved the best quality search and search agent products at every price point, we started working on latency." 00:31:50

Bootstrap a Data Flywheel by Competing Against Humans, Not Incumbents

Rather than trying to match Google's index on day one, Parallel launched an agent product that competed against human researchers — a dramatically lower bar — and used those real-world workflows to generate the empirical evals needed to build their actual search product. The lesson: find the adjacent market where your early-stage product is already better, generate real signal, then move up-market.

"We said it seems like humans sitting on search engines are far way easier to compete with than a search engine on day zero. So by building a product that was a search agent to do real work on top of web data, we were able to incrementally go build our index." 00:08:42

Optimize Your Docs for Agent Readers, Not Just Human Developers

Parallel tests their own documentation from the perspective of an AI agent reading it, not a human navigating a docs page. For any company whose customers build AI-native products, this is an immediately actionable shift in how to think about developer experience.

"For us the primary audience is an agent and that's how we test our docs." 00:50:17

Use Multiplier Thinking to Estimate Agentic Infrastructure Demand

When sizing a market or projecting infrastructure spend, don't use human query volume as the baseline. Use it as a starting point and multiply by the agent action chains each human intent triggers. A single background agent configured once generates ongoing query volume with no incremental human effort.

"One time I went and created one prompt to build this custom agent. Now it does tens and hundreds of web searches for every meeting I have. Every time I build a new agent for a new use case, that keeps multiplying." 00:30:36

6. Overlooked Insights

The Company Was Originally Incorporated as "Shapley Inc." — Revealing the True Core Thesis

This was mentioned in a single throwaway sentence but is enormously significant: the original name tells you that Parallel's deepest identity is not a search company or an AI company — it is a content economics company, built from first principles on a mathematical theory of fair value attribution. The search product is the delivery mechanism for the Shapley-based monetization thesis, not the other way around. Investors and competitors who focus only on the search infrastructure layer are missing what Parag himself considers the foundational insight.

"The company's original name was Shapley Inc. Really. When I incorporated... I was obsessing about everything to do with the problem space." 00:47:50

Public Company CEOs Should Optimize Earnings Reports for Agent Parsing — Right Now

Andrew Reed made a brief offhand observation that is immediately actionable and radically underappreciated: more people are consuming earnings transcripts through agents than are listening live or reading the transcript themselves. The implication is that how a CEO structures language in an earnings call — precision of numbers, clarity of forward guidance, named entities — now directly affects how agents summarize and propagate that information at scale. Companies that optimize IR communication for agent readability will have a structural advantage in how their story is understood and distributed.

"If I were a public company CEO today and I was doing an earnings report I would make very clear that the thing that I'm saying will be transcribed and interpreted correctly by the agents not just by people listening theoretically." 00:49:36